An agent that writes articles every day and connects them with internal links

Who the client was and what was getting in the way

A software company selling services, for which search visibility is one of the main sources of enquiries. The company did not need one article a month; it needed a steady flow of thematically consistent content to build authority in several focused areas. That means endless editorial work: someone has to maintain the topic plan, prevent articles from duplicating one another, and keep track of what should link to what.

Keyword selection was another challenge. In B2B services, it is easy to fall into the trap of answering questions such as “how much does it cost?” — queries where search engines often show a generated summary instead. Traffic from those searches does not reach the website, and even when it does, it rarely turns into a sales conversation. The content needed to target queries with commercial intent.

The third issue was speed. A new article needs to be submitted to search engines and connected to its topic cluster immediately. An article left on its own for two weeks is wasted effort. The fourth and most important requirement was that the agent strengthen content that was already close to ranking, rather than produce more isolated pages no one remembers.

What we built

We built an agent that manages the whole cycle, from choosing a topic to checking the results. Claude handles the editorial work; performance data comes from the Search Console API; and the agent’s state, article history, and relationship map are stored in Supabase. The process runs on a server on a schedule.

Each run starts from the company’s own knowledge graph. The agent checks which topic clusters already have coverage, where the gaps are, and what is actually earning impressions in each area. It chooses a topic based on that information, so the next article fills a gap instead of repeating what is already there.

Next comes keyword selection. The agent combines Search Console data with search suggestions and filters out queries that search engines answer without sending people to a website. For the chosen query, it reviews the existing search results and writes an article with a question-and-answer section marked up with structured data.

The article is published straight to production in two language versions. Immediately afterward, the agent submits it to search engines through IndexNow and starts building links. It does more than add links in the new article: it revisits older articles in the same cluster and adds links there to the new publication.

There is also a separate strengthening mode. The agent treats queries ranking between positions eight and thirty as opportunities to improve existing content, not as a reason to start a new article from scratch. After some time, it checks whether the changes worked.

How it works today

The agent runs continuously and publishes one or two articles a day in Polish and English. Within minutes, each new article is submitted to search engines and connected to the internal-link network in both directions: it links to older articles, and older articles link back to it.

The decision to write something new or improve an existing article is based on actual ranking data, not a plan drawn up a quarter earlier. No one at the company maintains a publication calendar; the order follows the gaps and the pages approaching a worthwhile ranking.

The outcome

The agent publishes one to two articles a day. Content is produced in two languages: Polish and English. Internal linking and search-engine submission happen automatically with every publication. The agent learns from Search Console data, so its next decisions are based on what actually ranks.

What we would do differently

The first version wrote good articles, and that was the problem: good articles without a topic cluster or internal links make little difference. We added the relationship network and strengthening mode later, and those are what produced the result. If starting again, we would build the cluster map and linking mechanism first, then connect the writing workflow.

We also underestimated how easily the agent drifts toward informational queries. Search engines answer these themselves, but the queries look attractive in search-volume statistics and tempt the model. The filter had to be strict and hard-coded; a gentle instruction in the prompt was not enough.